99 · A Hundred Model Masteries

This is how I would use AI first: find the ten LLMs from the biggest ecosystems, master one to the pinnacle, then move to the next, one by one. US: OpenAI's GPT, Anthropic's Claude, Google's Gemini, Meta's Llama, Microsoft's Phi, Cohere's Command, Mistral, AI21's Jurassic, Databricks' DBRX, Nvidia's Nemotron. India: Sarvam's OpenHathi, Krutrim, CoRover's BharatGPT, Zoho's Zia, Hanooman, Tech Mahindra's Project Indus, AI4Bharat's IndicBERT, JioHaptik. China: DeepSeek, Alibaba's Qwen, Baidu's Ernie, Tencent's Hunyuan, Zhipu's GLM, Moonshot's Kimi, MiniMax, ByteDance's Doubao, 01.AI's Yi, SenseNova. UK: DeepMind, Stability, PolyAI, Reka, Graphcore, Faculty's Frontier, Secondmind. And there are types of AI to learn for each use — marketing, customer acquisition, sales, video and image generation, shorts production. It's like recruiting. I'd brainstorm twenty hours a week, explore every feature of every model — do that for a hundred LLMs and you've functionally mastered a thousand platforms.

The first couple of months is just brainstorming twenty hours a day — chatting with every LLM, figuring out the best use of each, reading and building in sheer volume. The volume itself negates the quality question.

Cursor, Stealth, Nano Banana, OpenClaw, Lindy, Dex, Eleven Labs, Replicate, Perplexity Research, Openbase, TSMC-adjacent tooling — the AI stack to learn post-A-levels for the Manitoba months. Learning to drive these tools perfectly is a business multiplier on its own.

The smaller LLMs, and the open-source ones — the large Chinese models especially — are very liberal: speculative and creative in ways the US models aren't. Different tools, different temperaments; learn the personalities.

Interesting speculation on how the LLMs actually work under the data: there's input, integration, and output. Identify the public information the biggest firms release and feed the stack deliberately. We're still in the brute-force phase — train on everything — but I want the contractionary layer: intense training on data that matters. The counterargument: the data that matters is usually outliers, and outliers carry diminishing returns plus the risk of polluting the distribution. The rebuttal drags you into what LLMs fundamentally are: pattern recognition via input. Being selective about the data sources you feed your models is the key that most people skip. Before you build your own, know exactly what you're building on.

Learning how to code is kind of like learning how to build a car when you can just learn how to drive. In fact, nowadays even learning to drive is becoming adamant by the time our children grow up. The paradigm has shifted: the key is not building every skill by hand — it's identifying the list of skillsets necessitated, then using AI to feed them into the work while you hone judgment. Work in sheer volume, take action, and the way out reveals itself.

What if all the knowledge you accumulated about "how it must be done" is wrong now? What if it's just: find the skillsets needed, ask AI to be the expert, and figure it out together? Hold that question loosely — it's half true, and the half matters.